Identifying Site-specific Metastasis Genes and Functions

Identifying Site-specific Metastasis Genes and Functions G.P. GUPTA,* A.J. MINN,*§ Y. KANG,* ** P.M. SIEGEL,*# I. SERGANOVA,† C. CORDÓN-CARDO,‡ A.B. OLSHEN,¶ W.L. GERALD,‡ AND J. MASSAGUÉ* *Cancer Biology and Genetics Program and Howard Hughes Medical Institute, and Departments of †Neurology, ‡ Pathology, and ¶Biostatistics, Memorial Sloan-Kettering Cancer Center, New York, New York 10021 Metastasis is a multistep and multifunctional biological cascade that is the final and most life-threatening stage of cancer progression. Understanding the biological underpinnings of this complex process is of extreme clinical relevance and requires unbiased and comprehensive biological scrutiny. In recent years, we have utilized a xenograft model of breast cancer metastasis to discover genes that mediate organ-specific patterns of metastatic colonization. Examination of transcriptomic data from cohorts of primary breast cancers revealed a subset of site-specific metastasis genes that are selected for early in tumor progression. High expression of these genes predicts the propensity for lung metastasis independently of several classic markers of poor prognosis. These genes fulfill dual functions—enhanced primary tumorigenicity and augmented organ-specific metastatic activity. Other metastasis genes fulfill functions specialized for the microenvironment of the metastatic site and are consequently not selected for in primary tumors. These findings improve our understanding of metastatic progression, facilitate the interpretation of primary tumor gene expression data, and open several important possibilities for future clinical application. Tumorigenesis involves the temporal acquisition of genetic and epigenetic alterations that ultimately enable a cell to divide without concern for the homeostatic constraints that limit the growth of normal tissues. As such, cancer is not a static phenomenon, but rather a dynamic process that evolves over the course of tumor initiation and progression, and can manifest with an impressively diverse array of phenotypic properties from one primary tumor to the next. One of these properties is the ability of cancerous cells from one organ to invade and thrive at another organ. Also known as metastasis, distant recurrence is the leading cause for mortality in patients with solid tumors of most organs. However, not all primary tumors acquire the metastatic phenotype in the course of disease progression, and prospectively identifying which patients are most (or least) likely to develop metastases is of immense clinical importance. Furthermore, understanding the mechanisms that drive the formation of metastases may identify novel targets for much-needed therapy against this deadly biological process (see Fig. 1). From the perspective of an invasive cancer cell, not all potential sites for metastasis are created equal. Clinicians have observed for over a century that certain types of primary tumors are more likely to metastasize to specific organs (Fidler 2003). For example, advanced colon cancers frequently spread to liver, and breast cancers preferentially metastasize to bone and lung. On the basis of these clinical observations, Stephen Paget proposed in 1889 the “seed and soil hypothesis,” which postulated that tumor cells (the seeds) will only grow in a distant organ if they are competent to thrive in that microenvironment (the Present addresses: §Department of Radiation Oncology, University of Chicago, Chicago, Illinois 60637; **Department of Molecular Biology, Princeton University, Princeton, New Jersey 08544; #Departments of Medicine and Biochemistry, McGill University, Montreal, Quebec, Canada H34 1A4. soil). This theory, which placed prime emphasis on the cross talk between metastatic tumor cells and their microenvironment, was contested by James Ewing in 1926, when he proposed that metastatic propensities are dictated primarily by circulatory patterns—i.e., cells will metastasize to the organ to which they have the greatest vascular access. Subsequent analyses of patterns of metastatic spread in patients as well as in experimental models concluded that although regional recurrences were highly dependent on the efficiency of vascular perfusion, distant metastatic recurrence for most tumors was truly non-random, with no correlation to anatomically defined patterns of hematogenous or lymphatic circulation. Thus, Paget’s seed and soil hypothesis prevailed, although molecular determinants of this putative cross talk were still entirely unknown. For a cell to successfully metastasize to a distant organ, it must resist cell death pathways while accomplishing several distinct biological steps, including intravasation, adhesion, extravasation, angiogenesis, and growth in a foreign tissue (Chambers et al. 2002). Presumably, there are molecular mediators of these various processes, and the seed and soil hypothesis would suggest that at least some of these mediators are tissue-specific (Fidler 2003). Because of these complexities, metastasis is considered an inefficient process. In fact, it is postulated that malignant tumors release many thousands of cells into circulation daily, yet several orders of magnitude fewer metastases are ever observed in patients. Mouse models of experimental metastasis also recapitulate this phenomenon, where only a fraction of cancer cells are able to generate macroscopic metastases. Cancer biologists have harnessed this highly selective feature of metastasis to discover genes that are specifically enriched (or depleted) in those tumorigenic cells that give rise to distant metastases in animal models. How the multiple functions required for metastasis are selected for in the process of tumor progression is not Cold Spring Harbor Symposia on Quantitative Biology, Volume LXX. © 2005 Cold Spring Harbor Laboratory Press 0-87969-773-3. 149 150 GUPTA ET AL. Figure 1. Steps in tumor progression and metastatic dissemination and growth. A schematic depicting the various pathologically defined stages of tumor progression, as well as various functions associated with metastatic spread. The functions related to dissemination are suspected to be more general mediators of metastasis. Subsequent functions required for metastatic colonization (in purple box) may be unique to the microenvironments of the different metastatic sites. well understood, and is currently a subject for debate (Bernards and Weinberg 2002; Hynes 2003). Generally, it is agreed that the genomic and epigenomic instabilities known to exist in cancerous cells can spawn massive genetic heterogeneity within tumor cell populations. However, whether or not there are specific genetic events selected for during metastatic progression, and where and when this selection may be taking place, is still highly contentious. Conventional belief on this issue, largely from models of experimental metastasis, is that rare variants from primary tumor populations are selected for at the metastatic site based on an ability to survive and ultimately thrive at the distant site. However, why a cell with such unique abilities would be present with any prevalence in the primary tumor, coupled with the notion that metastasis itself is a physically inefficient process, makes it difficult to imagine why metastasis is not more rare than it actually is. A breakthrough in this conundrum was revealed when microarray analysis was conducted on primary tumors from breast cancer patients. It was discovered by several laboratories that genes expressed in the bulk primary tumor population were sufficient to predict whether a patient would develop distant metastatic recurrence (van’t Veer et al. 2002; Ramaswamy et al. 2003). These surprising findings demanded a reassessment of the prevailing models of metastasis. Some have interpreted from these results that primary tumors are, at a very early stage, destined to be either metastatic or non-metastatic, and that no further meaningful selection is necessary before cells from the primary tumor metastasize to a distant organ. However, whether any of the genes included in these poor prognosis signatures mediate metastasis remains an unanswered question. Additionally, the expression of poor prognosis genes in primary tumors does not explain the diversity of organ-specific metastasis patterns exhibited by advanced breast cancers. It has more recently been appreciated that poor prognosis signatures derived from different cohorts of patients by different laboratories yield distinct genes with very little overlap (van de Vijver et al. 2002; Jenssen and Hovig 2005; Wang et al. 2005). Thus, the biological meaning, as well as the clinical utility, of poor prognosis genes in primary tumors remains to be unraveled. In recent years, our laboratory has explored mechanisms of metastasis using a heterogeneous breast cancer cell line derived from the pleural effusion of a patient with widely metastatic breast cancer (Kang et al. 2003; Minn et al. 2005a,b). Through various techniques, we have identified subpopulations of cells within the parental cell line that exhibit distinct patterns of site-specific metastasis when inoculated into immunocompromised mice. We have demonstrated that these metastatic phenotypes can be linked to specific patterns of gene expression. By overexpressing candidate metastasis genes in weakly metastatic cells, or by knocking down their expression in aggressively metastatic cells, we have confirmed that many of these genes are mediators as well as markers of site-specific metastasis. Finally, we have applied these site-specific metastasis signatures to a cohort of primary breast cancers with known metastatic outcome, which has yielded significant insight into the biology driving this complex process. We present this body of work as a new methodological paradigm that couples experimental models of metastasis with the analysis of human breast tumors, in an attempt to discover clinically relevant mechanisms of breast cancer metastasis. GENETIC HETEROGENEITY DETERMINES DIFFERENCES IN METASTATIC POTENTIAL The MDA-MB-231 cell line is derived from the malignant pleural effusion of a patient with metastatic SITE-SPECIFIC METASTASIS GENES 151 Figure 2. Genetic heterogeneity determines metastatic phenotypes. (A) Multidimensional scaling plot of ~1200 genes that are differentially expressed among SCPs derived from the parental MDA-MB-231 cell line. (B) Metastatic phenotypes after intracardiac and tail vein inoculation of representative SCPs. (C) Expression of a Rosetta-like poor prognosis signature by several SCPs derived from MDA-MB-231 cells. The Pearson correlation coefficient between the different SCPs is invariably greater than 0.95. (D) Orthotopic inoculation of different tumor cell populations and subsequent surgical resection and monitoring for emergent lung metastases. The table shows lung metastatic activity for parental MDA-MB-231 cells, highly bone metastatic 1833 cells, moderately lung metastatic 1834 cells, and aggressively lung metastatic 4175 (LM2) cells. breast cancer (Cailleau et al. 1978). We postulated that this cell line might be composed of cells that are genetically heterogeneous. In fact, one may imagine that malignant cells in the pleural fluid of a patient with metastatic breast cancer may represent a demographic cross section of the circulating tumor cells that have been released from metastases in diverse organs. To test this hypothesis, we performed limiting dilution cloning of the parental cell line to derive several distinct populations of single cell-derived progeny, or SCPs. Transcriptomic analysis of these cells using Affymetrix HGU133A microarrays revealed that over 1200 genes were differentially expressed among these different populations (Minn et al. 2005b). By representing these gene expression differences in three dimensions using multidimensional scaling, three distinct subgroups of SCPs were identified (Fig. 2A). These findings confirmed that cells within the MDA-MB-231 cell line were genetically heterogeneous and exhibited distinct patterns of gene expression. We next sought to determine whether these genetic differences had implications for the metastatic potential of the different SCPs. Consequently, we xenografted the SCPs into either the left cardiac ventricle or the lateral tail vein of immunocompromised nude mice. To facilitate identification and monitoring of the emergent metastases, we engineered the SCPs to express a triple modality imaging vector encoding a fusion protein of thymidine kinase, GFP, and firefly luciferase (Ponomarev et al. 2004), and performed noninvasive bioluminescence imaging sequentially over time. To our surprise, the SCPs exhibited a diverse array of phenotypic patterns of metastatic spread (Fig. 2B). Some of the SCPs were robustly metastatic to the bones, yet displayed no metastatic activity to the lungs. Alternatively, some SCPs yielded aggressive metastases to the lungs and/or adrenal medulla, while being only mildly metastatic to the bones. A third group of SCPs exhibited dormant metastatic behavior, giving rise only to indolent growths that rarely developed into overt metastases. Gratifyingly, SCPs that were simi- 152 GUPTA ET AL. lar based on the multidimensional scaling plot of differentially expressed genes displayed similar metastatic behaviors. Thus, cells within the pleural effusion-derived MDA-MB-231 cell line were phenotypically diverse in their metastatic potential, and these differences correlated with distinct patterns of gene expression. ORGAN-SPECIFIC METASTATIC POTENTIAL IS NOT RELATED TO DIFFERENCES IN A POOR PROGNOSIS SIGNATURE By using supervised clustering of a cohort of primary breast cancers, van’t Veer and colleagues (van’t Veer et al. 2002) identified a 70-gene poor prognosis signature that classified breast cancers with a high likelihood of developing distant metastatic recurrence. This signature was validated on an independent cohort of 295 breast cancers and was shown to be an independent factor predicting patient prognosis (van de Vijver et al. 2002). We wanted to determine whether genes in this signature correlated with organ-specific patterns of metastasis. To this end, we confirmed that the parental MDA-MB-231 cell line expressed a Rosetta-type poor prognosis signature, containing 54 of the 70 poor prognosis genes identified by van’t Veer et al. that are represented on the Affymetrix HG-U133A microarray platform. This subset of genes was validated by two methods. First, these 54 genes performed nearly as well as the original 70-gene signature in predicting patient prognosis of the 78-tumor cohort from which the gene set was derived. Second, Affymetrix probe sets corresponding to these 54 genes were able to segregate a subgroup of patients with a worse prognosis in an independent cohort of primary breast cancers with at least 5 years of follow-up obtained at our institution. Parental MDA-MB-231 cells expressed this Rosetta-type poor prognosis signature in a manner similar to primary breast cancers that fell into the poor prognosis classification. In contrast, MCF-10A cells (an immortalized nontumorigenic mammary epithelial cell line) did not express the poor prognosis signature. All of the SCPs derived from MDA-MB-231 cells also expressed the poor prognosis signature. This was evident from a hierarchical clustering analysis of the SCPs combined with the MSKCC 82 primary breast cancer cohort. In addition, there was very little variation in the expression of the poor prognosis genes among the different SCPs (Fig. 2C, Pearson correlation coefficients greater than 0.95). Furthermore, none of the poor prognosis genes correlated with any of the organ-specific metastatic patterns exhibited by the different SCPs. Thus, although the poor prognosis genes may indicate whether a primary tumor is likely to develop distant metastatic recurrence, expression of these genes does not explain the diversity of metastatic patterns exhibited by advanced breast cancer cells. Interestingly, dormant SCPs that rarely gave rise to overt metastases also uniformly expressed the poor prognosis signature. This indicates that expression of these genes is not sufficient for metastasis and, consequently, that additional gene expression events must occur before cells gain a truly metastatic phenotype. GENES THAT MEDIATE SITE-SPECIFIC METASTASIS To identify genes that mediate organ-specific metastasis, we utilized in vivo selection of highly metastatic subpopulations from the weakly metastatic parental cell line by passaging it through immunocompromised mice. For bone metastasis assays we injected parental MDA-MB231 cells into the left cardiac ventricle, and for lung metastasis we introduced weakly metastatic parental or 1834 cells (a bone metastasis isolate from the parental cell line that did not exhibit any enrichment in metastatic activity) into the lateral tail vein. By extracting metastatic lesions and reinoculating them into mice to assay for enrichment in metastatic activity, we were able to isolate aggressively bone metastatic sublines after one round of in vivo selection (denoted BM1), and highly lung metastatic populations after two rounds of in vivo selection (denoted LM2). Transcriptomic analysis of these different in vivo selected subpopulations and the parental cell line enabled the elucidation of bone and lung metastasis gene expression signatures. The bone metastasis signature comprised 102 genes, of which 43 were overexpressed and 59 underexpressed in highly bone metastatic populations (Kang et al. 2003). Similarly, the lung metastasis signature contained 48 genes that were overexpressed and 47 genes that were underexpressed in aggressively lung metastatic LM2 populations (Minn et al. 2005a). Many of the genes in these signatures encoded secretory or cellsurface proteins, making them ideal candidates for enabling interactions between the metastatic tumor cells and their adopted microenvironments (see Fig. 3A,B). Interestingly, only 6 genes were concordantly expressed in both metastasis signatures. There are likely to be many genes that facilitate general metastatic activity, such as those that promote intravasation from the primary tumor into the circulation. However, our experimental approach was in principle seeking no such genes, but genes that mediate metastatic events in the distant organs (see purple box in Fig. 1). First, our starting cell line was already derived from cells that had escaped from the patient’s primary tumor and metastasized to the pleural cavity. In addition, our experimental metastasis assays involved direct inoculation into the circulation, thereby modeling the later steps of metastatic growth that are more likely to be site-specific. Nonetheless, the existence of distinct site-specific metastasis signatures expressed by subpopulations of cells from the same parental cell line confirmed the hypothesis that cells in malignant breast cancer pleural effusions are differentially genetically endowed to metastasize to various organs. To distinguish genes that mediate metastasis from those that simply correlate with and serve as markers for metastatic potential, we performed functional assays (Kang et al. 2003). Overexpression of interleukin-11 (IL11) and osteopontin (OPN) together, but neither alone, was sufficient to enhance the osteolytic bone metastatic activity of parental cells (Fig. 3C). Addition of either con- SITE-SPECIFIC METASTASIS GENES 153 Figure 3. Cooperation between functional mediators of tissue-specific metastasis. (A) Expression patterns of several overexpressed bone metastasis genes. Populations listed include the two replicates of the parental MDA-MB-231 cells, in vivo selected subpopulations, and SCPs. Metastatic activity of the different populations is color-coded. (B) Heatmap showing expression of several lung metastasis genes overexpressed in lung metastatic populations. Lung metastatic SCPs expressed only a subset of lung metastasis genes. (C) Summary of overexpression experiments demonstrating the cooperative action of different bone metastasis mediators. Relative bone metastatic strength was calculated from the time until 50% of the bone metastatic events occurred and the total percentage of mice that developed bone metastasis for each cohort, both of which were normalized to the basal metastatic activity of parental MDA-MB-231 cells. (D) Relative lung metastatic activity after transducing parental cells with various genes and gene combinations. Lung metastatic activity was determined by bioluminescent imaging after approximately 7 weeks post-xenografting, normalized to parental cells transduced with a vector control. nective tissue growth factor (CTGF) or chemokine (C-XC motif) receptor 4 (CXCR4) to these genes increased the rate and frequency of bone metastasis (Fig. 3C). These observations supported the notion that metastasis is a multistep and multifunctional process and, as such, multiple genes may be required to see an enhancement in the metastatic rate. Based on the previously known biology of these gene products, we postulated that CXCR4 may facilitate homing to and survival in the bone microenvironment, that IL-11 and OPN may cooperate in the recruitment and activation of host osteoclasts, and that CTGF may modify the extracellular environment to facilitate angiogenesis. In this manner, metastasis genes may fulfill different aspects of the cross talk between the tumor cells and stromal cells that are necessary to enable metastatic growth. Similar functional analysis identified nine lung metastasis genes that cooperated to promote aggressive lung metastatic growth (Minn et al. 2005a). Overexpression studies demonstrated that lung metastasis resulted from the cooperation of several extracellular modifiers, including the extracellular matrix molecule SPARC, the chemokine CXCL1, the mitogen Epiregulin, the matrix metalloproteinases MMP1 and MMP2, the cell surface receptors VCAM1 and IL13RA2, as well as intracellular modulators of gene expression and signaling including ID1 and COX2/PTGS2 (Fig. 3D). Importantly, overexpression of these genes had no effect on the bone metastatic activity of the parental cell line. RNAi-mediated knockdown of ID1, VCAM1, and IL13RA2 resulted in a greater than 10-fold reduction in lung metastatic growth within 6 weeks after tail vein injection. These findings confirmed that many of the genes selected for during in vivo selection of highly lung metastatic subpopulations were mediators of aggressive lung metastasis. Furthermore, these genes cooperated to facilitate lung-specific functions necessary for aggressive lung metastatic growth. 154 GUPTA ET AL. SITE-SPECIFIC METASTASIS GENES ARE EXPRESSED BY A SUBSET OF CELLS IN THE PARENTAL CELL POPULATION Although the site-specific metastasis signatures were generated by transcriptomic analysis of in vivo selected populations, they were also useful in identifying in vitro derived SCPs that were more or less metastatic to either bone or lung. Hierarchical clustering analysis of the SCPs with the bone metastasis signature identified a subgroup of SCPs that was genetically similar to the in vivo selected highly bone metastatic populations (Fig. 3A). Northern blot analysis of 46 SCPs from the parental MDA-MB-231 cell line for five of the most differentially expressed bone metastasis genes revealed SCPs that expressed a continuum of these genes, ranging from none of them to all five of them (Kang et al. 2003). The bone metastatic activity of these SCPs was directly correlated with the degree of expression of these genes, providing further evidence that these genes were mediators of osteolytic bone metastatic growth. By analyzing expression of bone metastasis genes in SCPs derived from in vivo selected populations, we noticed an approximately 5-fold enrichment in the proportion of cells expressing several bone metastasis genes. Thus, bone metastasis genes were expressed in a minority of cells in the malignant pleural effusion-derived parental cell line, and in vivo selection resulted in the enrichment of these preexisting highly bone metastatic cells. Similarly, the in vivo selected lung metastasis signature was also able to segregate lung metastatic SCPs from those that were not metastatic to the lungs (Minn et al. 2005a). When compared to the LM2 lung metastatic populations obtained through three rounds of in vivo selection, the lung metastatic SCPs only expressed a partial lung metastasis signature (Fig. 3B). In accordance with this observation, lung metastatic SCPs were approximately 10-fold less metastatic than the in vivo selected populations upon injection into the lateral tail vein. An observation of interest was that the genes in the lung metastasis signature were naturally divided into two categories. We postulated that genes that were expressed by both the lung metastatic SCPs and the LM2 populations may facilitate a baseline level of lung metastatic activity, which we describe as “lung metastagenicity.” Genes that were expressed exclusively by the most highly metastatic LM2 populations may confer functions enabling aggressive growth within the lung microenvironment, which we denote as “lung metastatic virulence.” Lung metastagenicity genes included ID1, COX2, CXCL1, MMP1, and many others. Examples of lung metastatic virulence genes were VCAM1, SPARC, IL13RA2, and MMP2. Genes from both of these categories were shown to be mediators of lung metastasis in functional assays. EXPRESSION OF SITE-SPECIFIC METASTASIS GENES IN PRIMARY BREAST CANCERS Because of the recent discovery of poor prognosis signatures in primary breast cancers, we wanted to examine whether our organ-specific metastasis signatures may also be expressed at this early stage. To ask this question, we utilized a cohort of 82 primary breast tumors with at least 3 years of follow-up, and with known organ-specific metastatic outcome (hereby referred to as the MSKCC cohort). Direct hierarchical clustering of all 82 primary tumors with the bone metastasis signature did not identify a subgroup of tumors that expressed the signature in a manner resembling the BM1 populations (Minn et al. 2005b). When the analysis was restricted only to patients that were known to develop metastatic recurrence, the bone metastasis signature segregated patients that went on to develop primarily bone metastasis from those that developed metastasis to other sites. Thus, the experimentally derived bone metastasis signature was only marginally expressed by primary breast tumors, and partial expression of this signature could not be used to prospectively identify patients with an increased likelihood of developing bone metastatic recurrence. Hierarchical clustering of this same primary tumor cohort with the lung metastasis signature yielded a dramatically different result (Minn et al. 2005a). A group of tumors in a highly reproducible branch of the dendrogram expressed the lung metastasis genes in a manner resembling the LM2 populations (Fig. 4A). This subset of tumors had molecular features of aggressive disease, including negative estrogen and progesterone receptor status, expression of a Rosetta-type poor prognosis signature, and a basaloid genotype. Examination of the metastatic outcome of these patients revealed a high prevalence of lung metastatic recurrence (Fig. 4A). This prompted us to examine these data in a statistically rigorous manner. First, univariate analysis of the lung metastasis signature identified expression of 12 genes that significantly correlated with tumors which gave rise to lung metastases. A classifier was generated by weighting the expression of lung metastasis genes according to the aforementioned univariate correlations, which identified patients that had a significantly higher likelihood of developing distant lung, but not bone, metastatic recurrence (p = 0.0018 for lung metastasis, and p = 0.31 [NS] for bone metastasis). In a separate analysis, weighting the expression of the lung metastasis genes according to the fold change exhibited by LM2 versus parental MDAMB-231 populations also created a classifier that distinguished a subgroup of patients with a high risk of developing lung metastasis. Thus, an experimentally derived lung metastasis signature was at least partially expressed by a subset of primary breast cancers, and these patients were more likely to develop lung metastases in the course of their disease. A biologically meaningful lung metastasis signature should be expressed by different cohorts of primary tumors transcriptomically profiled on different microarray platforms. We therefore validated our lung metastasis signature on the cohort of 78 primary breast cancers utilized by van’t Veer et al. using dual color Agilent cDNA microarrays (Minn et al. 2005a). Hierarchical clustering revealed a group of tumors that coexpressed the lung metastasis signature in a manner resembling the LM2 in vivo selected populations. Of note, the 9 functionally validated lung metastasis genes were remarkably overex- SITE-SPECIFIC METASTASIS GENES 155 Figure 4. Expression of lung metastasis genes by primary breast cancers. (A) Unsupervised hierarchical clustering of the MSKCC cohort (82 tumors) with the 12 most univariately significant lung metastasis genes (correlation with lung metastatic outcome, p<0.05). Also shown are a Rosetta-like poor prognosis signature, ER, PR, and Her2 expression, as well as keratin markers of basal and luminal subtypes of breast cancer. Patients identified with a red dot developed metastases to sites other than the lung, whereas patients labeled with a black dot suffered from lung metastases. Highlighted in a light blue background is a robust branch of the dendrogram containing tumors that expressed the lung metastasis genes in a manner resembling the high lung metastatic populations derived from MDA-MB-231 cells. This group of tumors was enriched for those with a high likelihood of developing lung metastatic recurrence. (B) Lung (top) and bone (bottom) metastasis-free survival curves for patients that expressed the lung metastasis signature (red) and those that did not express the lung metastasis signature. pressed in this group of tumors. Although organ-specific metastatic outcome is not publicly available for this cohort, time until distant metastatic recurrence and overall survival of these patients is known. The majority of primary breast cancers that expressed the lung metastasis signature in this cohort went on to develop distant metastatic recurrence and had a poor overall survival. We hypothesize that many of these patients may have suffered from lung metastasis. Interestingly, not all of the genes in the experimentally derived lung metastasis signature were informative in predicting primary tumors that went on to develop lung metastasis. For this reason, training the signature on a primary tumor cohort to weight genes according to their robustness of expression among tumors that coexpress lung metastasis genes may engender a more accurate algorithm that is better suited to classify primary breast cancers. Because clinical outcome is not needed for algorithm training, the cohort studied by van’t Veer et al. was used to generate a classifier that accurately segregated tumors which most resembled the LM2 populations (Minn et al. 2005a). Six of the nine functionally validated genes were among the most heavily weighted genes in this classifier. The remaining three genes were SPARC, MMP2, and IL13RA2, which were all lung metastasis virulence genes in MDA-MB-231 cells. This trained classifier was the most accurate predictor of lung metastatic recurrence in the MSKCC cohort (Fig. 4B), providing further evi- dence that the lung metastasis signature is biologically relevant, reproducible, and informative in identifying breast cancer patients with a high likelihood of developing lung metastatic recurrence. A SUBSET OF LUNG METASTASIS GENES, BUT NOT BONE METASTASIS GENES, FACILITATE PRIMARY TUMORIGENICITY The selective pressure that encourages expression of metastasis genes in a subset of primary tumors is not apparent. One possibility is that some metastasis genes may facilitate growth of the primary tumor, thereby increasing the representation of cells expressing these genes in the cancerous population. This is supported by the clinical observation that larger breast cancers tend to have a poorer prognosis, with an increased likelihood of metastatic recurrence. According to this postulate, lung metastasis genes that were informatively expressed in primary tumors might encourage primary tumor growth, whereas lung metastasis genes that were not predictive in primary tumors, and all bone metastasis genes, might have no effect on primary tumorigenicity. To test this hypothesis, we established an orthotopic model that mimicked breast cancer progression in patients (Fig. 2D). Injection of MDA-MB-231 populations in the mouse mammary fat pad gave rise to primary tumors. When these tumors were surgically resected, lung 156 GUPTA ET AL. Figure 5. Lung metastagenicity and primary tumorigenicity. (A) Relative tumor growth rates of different MDA-MB-231 populations after orthotopic implantation into the mouse mammary fat pad. (B) Tumor growth rates after reducing the expression of various lung metastasis genes in the 4175 (LM2) population. (C) Model depicting different subsets of lung metastasis genes and the roles they play in primary tumorigenicity, lung metastagenicity, and lung metastatic virulence. Bone metastasis genes did not significantly affect primary tumor growth, and as such only have a role in bone metastatic growth. metastases could be detected only in mice injected with cells that were selected for aggressive lung metastatic activity (Fig. 2D). We also noticed a difference in primary tumor growth rate upon orthotopic implantation of these populations. Although tumors formed by parental and bone metastatic populations grew at a similar rate in the mammary fat pad, the tumors formed by lung metastatic cell populations grew more rapidly in a manner proportional to their lung metastatic aggressiveness (Fig. 5A). This increase in primary tumor growth rate was not due to differences in the proliferative rate of the cells, because immunohistochemical analysis of cell division did not reveal significant disparities. Consequently, genes in the lung metastasis signature facilitated primary tumorigenicity, whereas bone metastasis genes had no effect on primary tumor growth rate. To identify which of the lung metastasis genes were also facilitating primary tumorigenicity, we performed orthotopic injections of LM2 cells expressing RNAi vectors targeting SPARC, IL13RA2, VCAM1, or ID1. Whereas targeted inhibition of all of these genes had an effect on inhibiting lung metastasis both via tail vein assays and after surgical resection of primary mammary fat pad tumors, only ID1 had a role in promoting primary tumorigenicity (Fig. 5B). Interestingly, ID1 was the only lung metastagenicity gene among the four genes tested, and was among the 18 most heavily weighted genes in the lung metastasis classifier of primary breast cancers. Thus, lung metastasis genes can be divided into two categories: those that facilitate both primary tumor formation and basal lung metastatic activity, and those that enable aggressive growth in the lung microenvironment without affecting growth in the primary tumor (Fig. 5C). LESSONS LEARNED By xenografting a malignant pleural effusion-derived breast cancer cell line into immunocompromised mice, we have learned several valuable lessons regarding principles governing metastasis. First, metastatic cells from the same primary tumor exhibit diverse metastatic tropisms, reflecting the genetic heterogeneity that pervades metastatic breast cancer populations. Second, genes that mediate metastasis to lung and bone are distinct, and encode many cell surface and secreted proteins that enable cross talk between the metastatic tumor cells and these two divergent organ microenvironments. Together, these gene products act in concert to facilitate the multifunctional task of metastasis. The recent findings of others have identified poor prognosis genes that, when expressed in primary tumors, predict a poor prognosis for breast cancer patients. The organ-specific metastasis genes that we have identified are distinct from these poor prognosis signatures. In fact, analysis of SCPs from MDA-MB-231 cells revealed that expression of these poor prognosis genes is not sufficient SITE-SPECIFIC METASTASIS GENES to enable metastasis. Rather, malignant breast cancer cells become overtly metastatic only when organ-specific metastasis signatures are layered upon this predisposing genetic platform. Analysis of organ-specific metastasis signatures in primary breast carcinomas has provided substantial biological insight into the mechanisms driving metastasis. Although bone metastasis genes were only weakly expressed in primary tumors that later disseminated to the bones, the lung metastasis signature was significantly expressed by a subset of breast cancers with a high likelihood of metastasizing to the lungs. This difference may be due to partial similarities between the breast and lung microenvironments that do not exist between breast and bone. In fact, cells expressing lung, but not bone, metastasis genes are significantly more tumorigenic when orthotopically inoculated into the mammary fat pad of immunocompromised mice. By selectively targeting different genetic mediators of lung metastasis, we exposed two distinct types of lung metastasis genes; one group that facilitated both primary tumorigenicity and lung metastatic ability, and another group that mediated aggressive growth exclusively in the lung microenvironment. Because none of the bone metastasis genes enhanced primary tumorigenicity, they are all examples of the latter category of metastasis genes (Fig. 5C). FUTURE DIRECTIONS Microarray analysis has unveiled a new era in the understanding of breast cancer. Examination of global differences in gene expression has confirmed the existence of distinct molecular subtypes of breast cancer with different clinical tendencies (Perou et al. 2000; Sorlie et al. 2003), a long-suspected postulate of breast cancer pathologists and clinicians. Microarray technology has also enabled the identification of gene expression signatures that indicate a poor prognosis for breast cancer patients (van de Vijver et al. 2002; Wang et al. 2005). However, the clinical utility of this application is hampered by concerns regarding the reproducibility and robustness of the correlations between these poor prognosis genes and clinical outcome, especially when tested on different patient populations using different microarray platforms (Jenssen and Hovig 2005). Indeed, clinical trials are currently ongoing to address some of these issues. Nonetheless, cancer biologists have, as yet, been unable to make a mechanistic link between the expression of these poor prognosis genes and the eventual clinical outcome that afflicts breast cancer patients. Consequently, improvements in the design of microarray-based experiments are necessary before the wealth of information provided by microarray analyses is harnessed for its true potential. One way to extract biological meaning from microarrays of clinical samples may be to use an experimentally tractable model of cancer as a biological filter for the identification of functionally relevant metastasis genes. For example, a recent analysis of a “wound signature,” derived by identifying serum-responsive genes in fibroblasts cultured in vitro, revealed that it classified tumors 157 into good and poor prognosis categories almost as effectively as the van’t Veer et al. poor prognosis signature (Chang et al. 2004, 2005). In contrast to the van’t Veer et al. signature, the wound signature is based on biological principles and excites intriguing future experimental pursuits using both model systems and clinical specimens. Our organ-specific metastasis signatures adopted the paradigm of functional derivation in a model system and subsequent validation using clinical samples. For this reason, we anticipate that the clinical correlations of the lung metastasis genes with lung metastatic outcome may be more reproducible across patient cohorts. Indeed, we have observed a subgroup of breast cancers expressing the lung metastasis signature in several other publicly available gene expression data sets derived from different microarray platforms. Whether these patients were also predisposed for developing lung metastasis is currently being investigated. In addition, we are developing quantitative RT-PCR methodology for the analysis of paraffin-extracted RNA (Cronin et al. 2004; Paik et al. 2004), which would enhance the validation potential of this signature as a clinically useful prognostic assay. It is our hope that because this signature is based on several experimentally validated mediators of lung metastasis, it may be more reproducible than other correlation-based gene expression signatures. Our experimental approach could be applied to identify genes mediating metastasis to other clinically important sites such as the brain or the liver, or metastasis by other types of primary tumors. Furthermore, organ-specific metastasis signatures may have implications for cancer management and therapy. The prospective identification of subgroups of patients coexpressing sets of genes that collectively facilitate metastasis may aid in the selection of appropriate patient populations for the administration of novel or preexisting metastasis therapies. For example, patients with primary tumors that express the lung metastasis signature may be monitored more frequently by thoracic computed tomography (CT) scans, and treated more aggressively with conventional chemotherapies. Additionally, experimental drugs inhibiting specific biological pathways may be tested in a combinatorial fashion selectively in this high-risk group of patients. It is not difficult to imagine that targeted therapies which proved unsuccessful as single agents may be efficacious when rationally combined with other drugs. A thorough understanding of the biological mechanisms by which metastasis genes facilitate tumorigenicity and/or metastasis should drive the rational design of clinical trials to test new therapeutic strategies aimed at selectively targeting metastases, while being nontoxic to the patient. Metastatic cancer remains, for the most part, a complex and incurable disease. In recent years, cancer biologists have accumulated a daunting array of technologies that facilitate the modeling, mechanistic dissection, and gene discovery of cancer progression. The next era in the fight against cancer has already begun, and it is founded upon the interdisciplinary efforts of engineers, biologists, biostatisticians, computer scientists, and clinicians. We are 158 GUPTA ET AL. optimistic that this interdisciplinary communication will make the coming years among the most exciting in the war against cancer. REFERENCES Bernards R. and Weinberg R.A. 2002. A progression puzzle. Nature 418: 823. Cailleau R., Olive M., and Cruciger Q.V. 1978. Long-term human breast carcinoma cell lines of metastatic origin: Preliminary characterization. In Vitro 14: 911. Chambers A.F., Groom A.C., and MacDonald I.C. 2002. Dissemination and growth of cancer cells in metastatic sites. Nat. Rev. Cancer 2: 563. Chang H.Y., Sneddon J.B., Alizadeh A.A., Sood R., West R.B., Montgomery K., Chi J.T., van de Rijn M., Botstein D., and Brown P.O. 2004. Gene expression signature of fibroblast serum response predicts human cancer progression: Similarities between tumors and wounds. PLoS Biol 2: E7. Chang H.Y., Nuyten D.S., Sneddon J.B., Hastie T., Tibshirani R., Sorlie T., Dai H., He Y.D., van’t Veer L.J., Bartelink H., van de Rijn M., Brown P.O., and van de Vijver M.J. 2005. Robustness, scalability, and integration of a wound-response gene expression signature in predicting breast cancer survival. Proc. Natl. Acad. Sci. 102: 3738. Cronin M., Pho M., Dutta D., Stephans J.C., Shak S., Kiefer M.C., Esteban J.M., and Baker J.B. 2004. Measurement of gene expression in archival paraffin-embedded tissues: Development and performance of a 92-gene reverse transcriptase-polymerase chain reaction assay. Am. J. Pathol. 164: 35. Fidler I.J. 2003. The pathogenesis of cancer metastasis: The ‘seed and soil’ hypothesis revisited. Nat. Rev. Cancer 3: 453. Hynes R.O. 2003. Metastatic potential: Generic predisposition of the primary tumor or rare, metastatic variants-or both? Cell 113: 821. Jenssen T.K. and Hovig E. 2005. Gene-expression profiling in breast cancer. Lancet 365: 634. Kang Y., Siegel P.M., Shu W., Drobnjak M., Kakonen S.M., Cordon-Cardo C., Guise T.A., and Massagué J. 2003. A multigenic program mediating breast cancer metastasis to bone. Cancer Cell 3: 537. Minn A.J., Gupta G.P., Siegel P.M., Bos P.D., Shu W., Giri D.D., Viale A., Olshen A.B., Gerald W.L., and Massagué J. 2005a. Genes that mediate breast cancer metastasis to lung. Nature 436: 518. Minn A.J., Kang Y., Serganova I., Gupta G.P., Giri D.D., Doubrovin M., Ponomarev V., Gerald W.L., Blasberg R., and Massagué J. 2005b. Distinct organ-specific metastatic potential of individual breast cancer cells and primary tumors. J. Clin. Invest. 115: 44. Paik S., Shak S., Tang G., Kim C., Baker J., Cronin M., Baehner F.L., Walker M.G., Watson D., Park T., Hiller W., Fisher E.R., Wickerham D.L., Bryant J., and Wolmark N. 2004. A multigene assay to predict recurrence of tamoxifen-treated, node-negative breast cancer. N. Engl. J. Med. 351: 2817. Perou C.M., Sorlie T., Eisen M.B., van de Rijn M., Jeffrey S.S., Rees C.A., Pollack J.R., Ross D.T., Johnsen H., Akslen L.A., Fluge O., Pergamenschikov A., Williams C., Zhu S.X., Lonning P.E., Borresen-Dale A.L., Brown P.O., and Botstein D. 2000. Molecular portraits of human breast tumours. Nature 406: 747. Ponomarev V., Doubrovin M., Serganova I., Vider J., Shavrin A., Beresten T., Ivanova A., Ageyeva L., Tourkova V., Balatoni J., Bornmann W., Blasberg R., and Gelovani Tjuvajev J. 2004. A novel triple-modality reporter gene for whole-body fluorescent, bioluminescent, and nuclear noninvasive imaging. Eur. J. Nucl. Med. Mol. Imaging 31: 740. Ramaswamy S., Ross K.N., Lander E.S., and Golub T.R. 2003. A molecular signature of metastasis in primary solid tumors. Nat. Genet. 33: 49. Sorlie T., Tibshirani R., Parker J., Hastie T., Marron J.S., Nobel A., Deng S., Johnsen H., Pesich R., Geisler S., Demeter J., Perou C.M., Lonning P.E., Brown P.O., Borresen-Dale A.L., and Botstein D. 2003. Repeated observation of breast tumor subtypes in independent gene expression data sets. Proc. Natl. Acad. Sci. 100: 8418. van de Vijver M.J., He Y.D., van’t Veer L.J., Dai H., Hart A.A., Voskuil D.W., Schreiber G.J., Peterse J.L., Roberts C., Marton M.J., Parrish M., Atsma D., Witteveen A., Glas A., Delahaye L., van der Velde T., Bartelink H., Rodenhuis S., Rutgers E.T., Friend S.H., and Bernards R. 2002. A gene-expression signature as a predictor of survival in breast cancer. N. Engl. J. Med. 347: 1999. van’t Veer L.J., Dai H., van de Vijver M.J., He Y.D., Hart A.A., Mao M., Peterse H.L., van der Kooy K., Marton M.J., Witteveen A.T., Schreiber G.J., Kerkhoven R.M., Roberts C., Linsley P.S., Bernards R., and Friend S.H. 2002. Gene expression profiling predicts clinical outcome of breast cancer. Nature 415: 530. Wang Y., Klijn J.G., Zhang Y., Sieuwerts A.M., Look M.P., Yang F., Talantov D., Timmermans M., Meijer-van Gelder M.E., Yu J., Jatkoe T., Berns E.M., Atkins D., and Foekens J.A. 2005. Gene-expression profiles to predict distant metastasis of lymph-node-negative primary breast cancer. Lancet 365: 671.